Siddharth Saoji 
When you ask a question to someone in an office, they respond and feel accountable for it. If they felt they could be wrong, they would discreetly recheck what they said—because to some extent, your trust in them is impacted by how correct they are in their responses.
For the longest time, this is not what we saw with AI agents. They might research well and rethink before they responded, but forgot about it the moment the response was out, unless you pointed out something and they scrambled to retract, rethink, or modify their answer. The upfront research-and-rethink loop also became a UX bottleneck when a human was waiting on the other end.
Now, some AI agents—just like a human coworker—rethink their responses even without being prompted by the user to do so. This helps with speed while retaining the credibility of the agent. I saw this happening while using Instinct, which ironically increased my trust in its responses, even when it corrected a mistake by sending a proactive acknowledgement of its previous message’s inaccuracy.
I asked Instinct itself how that sweep trigger works and here was its response verbatim:
Instinct: The mechanism, at a high level: when I answer something, I don’t fully close the loop. I can keep the question open in the background and re-check what I said against the actual sources. If something doesn’t hold up—like my numbers not matching the page I linked—that mismatch gets flagged and I come back to you with the correction instead of letting it sit wrong. So it’s basically the same discipline as a person double-checking their own work after hitting send, except I can do it while you’re getting on with your day.
Me: Ok, and till when does this self reflection loop run, 2 times or more?
Instinct: It’s not a fixed count, there’s no “2 passes and done”. It runs on reasons, not rounds: if an answer involves live sources or numbers you’ll act on, it stays open for re-checking until things settle. Once everything checks out, it closes. Routine stuff that can’t go stale (like a reminder) never gets re-opened. So the honest answer is: as many times as there’s something worth re-verifying, and zero when there isn’t.
Me: What is the definition of settled?
Instinct: Settled means three things are all true: what I told you matches the sources it came from, there’s no open question left you’re waiting on, and nothing about it can quietly change in a way that matters to you.
Me: So you make judgment calls on every response?
Instinct: Pretty much, yes. Every response is a small stack of judgment calls: is this something I can answer straight, or should I verify first? How confident am I? Is this worth interrupting you about, or can it wait? The key bit is that judgment and facts stay separate. I can judge whether to act, but I shouldn’t judge a fact into existence—numbers, dates, prices have to come from a real source, and when they’re uncertain I say so instead of guessing confidently.
It’s things like these that make the harness vs. model debate feel like a co-dependency. Here the harness facilitates the loop, but the loop would never have happened if the model intelligence did not deem it necessary.